Chapter 10 Social Search Social search describes search acts that make use of social interactions with others. These interactions may be explicit or implicit, co-located or remote, synchronous or asynchronous Social search Search within a social environment Communities of users actively participating in the search process Goes beyond classical search tasks Facilitates the information seeking process [Evans 08] 2 [Evans 08] Evans et al. Towards a Model of Understanding Social Search. In Proc. of Conf. on Computer Supported Cooperative Work Social vs. Standard Search Key differences Users interact with the system (standard & Social) Users interact with one another in a open/social environment implicitly (reading)/explicitly (writing) such as Visiting social media sites, e.g., YouTube Browsing through social networking sites, e.g., Facebook 3 Web 2.0 Social search includes, but is not limited to, the so- called social media site Collectively referred to as Web 2.0 as opposed to the classical notion of the Web (Web 1.0) Social media sites User generated content Users can tag their own and others content Users can share favorites, tags, etc., with others Provide unique data resources for search engines Example. YouTube, MySpace, Facebook, LinkedIn, Digg, Twitter, Flickr, Del.icio.us, and CiteULike 4 Social Media/Networking Sites 5 SM Services Operational/Functional Dimensions Level of Collaboration Content: Audio Content: Video Content: Image Content: Text Content: Aggregation Provider Censorship User CensorshipPrivacyCommunication TypeProvides API FacebookHIGHNONELOWMEDIUMHIGHLOWNONEHIGHBOTHMANY-TO-MANYGOOD LinkedInLOWNONE LOWHIGHNONE HIGHBOTH1-TO-MANYNO TwitterMEDIUMNONE HIGH NONEHIGHBOTHMANY-TO-MANYGOOD DeliciousLOWNONE LOWHIGHNONE PUBLICMANY-TO-MANYGOOD FlickrMEDIUMNONEHIGH LOWNONE HIGHBOTHMANY-TO-MANYGOOD YouTubeHIGHNONEHIGHNONELOWNONELOWHIGHBOTHMANY-TO-MANYGOOD SkypeLOWHIGH NONELOWNONE HIGHPRIVATE1-TO-1NO Last.fmLOWHIGHNONE LOWNONELOWBOTHMANY-TO-MANYGOOD yelp.comLOWNONE HIGH NONE 1-TO-MANYOK WikiAnswersHIGHNONE LOWHIGHNONE HIGHPUBLICMANY-TO-MANY World of WarcraftMEDIUM HIGHNONELOWNONE LOWBOTHMANY-TO-MANYNO Social Media/Network Sites 6 FacebookLinkedInDeliciousTwitterLibraryThingFlickrYouTubeMySpaceLast.fmCiteULike Web Searchyes maybe yesmaybe Yes (e.g., Google Scholar) Recommendation friends, groups, ads, links, connections, jobs, ads, books, news art, links, groups books, links, news artciles, movies, videos, groups people, location books, groups, linksusers, pics, travel, groups, links, videos video, links, people, friends, groups, ads, links, (provides rec based on search habits) friends, music, concerts (the rec application exists) artciles, links, groups Filtering messages, activity, jobs, news art, links links, news articles, Messages (hashtag) conversations, groups picsvideosmessages, activity, messages, music new papers, articles Ads Suggestion yes maybe yes Collaborative Searching/Filtering yesdepending on topic yes depending on topicyes maybe depending on topic User Similarity (profile) yes maybeyes maybe yes Personal Interest Identification yesdepending on domain yes yes (but may not be comprehensive) yes depending on topic Topic Identification maybe no yes No (or not really comprehensive) maybe yes (within area of study) Tag (matching/ suggestions) no yesmaybeyes noyes Social Search Topics Online user-interactive data, which provide a new and interesting search experience User tags: users assign tags to data items, a manual indexing approach Searching within communities: virtual groups of online users, who share common interests, interact socially, such as blogs and QA systems Recommender systems: individual users are represented by their profiles (fixed queries long-term info. need) such as CNN Alert Service, Amazon.com, etc. Peer-to-peer network: querying a community of nodes (individual/organization/search engine) for an info. need Metasearch: a special case of P2P all the nodes are SEs 7 User Tags and Manual Indexing Then: Library card catalogs Indexing terms chosen with search in mind Experts generate indexing terms manually Terms are very high quality based on the US Library of Congress (LOC) Subject Headings standardized by the LOC Terms chosen from controlled/fixed vocabulary and subject guides (a drawback) Now: Social media tagging Social media sites allow users to generate own tags manually (+) Tags not always chosen with search in mind (-) Tags can be noisy or even incorrect and without quality control (-) Tags chosen from folksonomies, user-generated taxonomies (+) 8 Social Search Topics Example. Some of the 116 million tags of LibraryThing, which archives 95 million book records w/ 1.93 million users (04/15) 9 Social Tagging According to [Guan 10] Social tagging services allow users to annotate online resources with freely chosen keywords Tags are collectively contributed by users and represent their comprehension of resources. Tags provide meaningful descriptors of resources and implicitly reflect users interests Tagging services provide keyword- based search, which returns resources annotated by given tags 10 [Guan 10] Guan et al. Document Recommendation in Social Tagging Services. In Proc. of Intl. Conf. on World Wide Web. 2010 Types of User Tags Content-based Tags describe the content of an item, e.g., car, woman, sky Context-based Tags describe the context of an item, e.g., NYC, empire bldg Attribute-based Tags describe the attributes of an item, e.g., Nikon (type of camera), black and white (type of movie), etc. Subjective-based Tags subjectively describe an item, e.g., pretty, amazing, etc. Organizational-based Tags that organize items, e.g., to do, not read, my pictures, 11 Searching Tags Searching collaboratively tagged items, i.e., user tags, is challenging Most items have only a few tags, i.e., complex items are sparely represented, e.g., aquariums goldfish, which is the vocabulary mismatch problem Tags are very short Boolean (AND/OR), probabilistic, vector space, and language modeling will fail if use navely High precision but low recall for conjunctive (AND) queries Low precision but high recall for disjunctive (OR) queries 12 Tag Expansion Can overcome vocabulary mismatch problem, such as aquariums and topical fish, by expanding tag representation with external knowledge Possible external sources Thesaurus Web search results Query logs After tags have been expanded, can use standard retrieval models 13 Tag Expansion Using Search Results 14 Age of Aquariums - Tropical Fish Huge educational aquarium site for tropical fish hobbyists, promoting responsible fish keeping internationally since The Krib (Aquaria and Tropical Fish) This site contains information about tropical fish aquariums, including archived usenet postings anddiscussions, along with new... Keeping Tropical Fish and Goldfish in Aquariums, Fish Bowls, and... Keeping Tropical Fish and Goldfish in Aquariums, Fish Bowls, and Ponds at AquariumFish.net. fish tropical aquariums goldfish bowls P(w | tropical fish) Example. Web search results enhance a tag representation, tropical fish, a query A retrieved snippet Pseudo-relevance feedback over related terms Searching Tags Even with tag expansion, searching tags is challenging Tags are inherently noisy (off topic, inappropriate) and incorrect (misspelled, spam) Many items may not even be tagged, which become virtually invisible to any search engine Typically easier to find popular items with many tags than less popular items with few/no tags How can we automatically tag items with few or no tags? Uses inferred tags to Improve tag search Automatic tag suggestion 15 Methods for Inferring Tags TF-IDF: wt(w) = log(f w,D + 1) log(N / df w ) Suggest tags that have a high TF-IDF weight in the item Only works for textual items Classification (determines the appropriateness of a tag) Train binary classifier for each tag, e.g., using SVM Performs well for popular tags, but not as well for rare tags Maximal marginal relevance Finds relevant tags to the item and novel with respect to others where Sim item (t, i) is the similarity between tag t and item i, i.e., T i Sim tag (t i, t) is the similarity between tags t i and t ( 0..1), a tunable parameter 16 Large, if t is very relevant to T i, but differs from other tags of T i Using TF/IDF Using query results Browsing and Tag Clouds Search is useful for finding items of interest Browsing is more useful for exploring collections of tagged items Various ways to visualize collections of tags Tag clouds (show the popularity of tags based on sizes) (Tags are) Alphabetically order and/or weighted Formatted/sorted according to popularity 17 Sample Tag Cloud 18 animals architecture art australia autumn baby band barcelona beach berlin birthday black blackandwhite blue california cameraphone canada canon car cat chicago china christmas church city clouds color concert day dog england europe family festival film florida flower flowers food france friends fun garden germany girl graffiti green halloween hawaii holiday home house india ireland italy japan july kids lake landscape light live london macro me mexico music nature new newyork night nikon nyc ocean paris park party people portrait red river rock sanfrancisco scotland sea seattle show sky snow spain spring street summer sunset taiwan texas thailand tokyo toronto travel tree trees trip uk usa vacation washington water wedding Tag Clouds As defined in [Schrammel 09], tag clouds are Visual displays of set of words (tags) in which attributes of the text such as size, color, font weight, or intensity are used to represent relevant properties, e.g., frequency of documents linked to the tag A good visualization technique to communicate an overall picture 19 [Scharammel 09] Schrammel et al. Semantically Structured Tag Clouds: an Empirical Evaluation of Clustered Presentation Approaches. In Proc. of Intl Conf on Human Factors in Computing Systems Searching with Communities What is an online community? Groups of entities (i.e., users, organizations, websites) that interact in an online environment to share common goals, interests, or traits Besides tagging, community users also post to newsgroups, blogs, and other forums To improve the overall user experiments, web search engines should automatically find the communities of a user Example. Baseball fan community, digital photography community, etc. Not all communities are made up of humans! Web communities are collections of web pages that are all about a common topic 20 Online Communities According to [Seo 09] Online communities are valuable information sources where knowledge is accumulated by interactions between people Online community pages have many unique textual or structural features, e.g., A forum has several sub-forums covering high-level topic categories Each sub-forum has many threads A thread is a more focused topic-centric discussion unit and is composed of posts created by community members 21 [Seo 09] Seo et al. Online Community Search Using Thread Structure. In Proc. of ACM Conf. on Information & Knowledge Management Finding Communities How to design general-purpose algorithms for finding every possible type of on-line community? What are the criteria used for finding a community? Entities (users) within a community are similar to each other Members of a community are likely to interact more with one another of the community than those outside of the community Can represent interactions between a set of entities as a graph Vertices (V) are entities Edges (E), directed or undirected, denote interactions of entities Undirected edges represent symmetric relationships Directed edges represent non-symmetric or causal relationships 22 HITS Hyperlink-Induced Topic Search (HITS) algorithm can be used to find communities A link analysis algorithm, like PageRank Each entity has a hub and authority score Based on a circular set of assumptions Good hubs point to good authorities Good authorities are pointed to by good hubs Iterative algorithm: 23 Authority score of p is the sum of the hub scores of the entities pointing at p Hub score of p is the sum of the authority scores pointed at by p Form community (C) Apply the entity interaction graph to find communities Identify a subset of the entities (V), called candidate entities, be members of C (based on common interest) Entities with large authority scores are the core or authoritative members of C to be a strong authority, an entity must have many incoming edges, all with moderate/large hub scores, and To be a strong hub, an entity must have many outgoing edges, all with moderate/large authority scores Vertices not connected with others have hub and authority scores of 0 24 HITS Finding Communities Clustering Community finding is an inherently unsupervised learning problem Agglomerative or K-means clustering approaches can be applied to entity interaction graph to find communities Use the vector representation to capture the connectivity of various entities Compute the authority values based on the Euclidean distance Evaluating community finding algorithms is hard Can use communities in various ways to improve web search, browsing, expert finding, recommendation, etc. 25 Graph Representation Node: Vector: Question Answering Goal Automatically answer questions submitted by humans in a natural language form Approaches Rely on techniques from diverse areas of study, e.g., IR, NLP, Onto, and ML, to identify users info. needs & textual phrases potentially suitable answers for users Exploit Data from Community Question Answering Systems (CQA) (Web) Data Sources, i.e., doc corpus 27 Question Answering (QA) n Question answering (QA) is a specialized form of IR n Given a collection of documents/collaborative QA system, the QA system attempts to retrieve correct answers to questions posted in natural language n Unlike search engines, QA systems generate answers instead of providing ranked lists of documents n Current (non-collaborative) QA systems extract answers from large corpora such as the Web n Fact-based QA limits range of informational questions to those with simple, short answers who, where, why, what, when, how (5W 1H/WH) questions 28 CQA-based approaches Analyze questions (& corresponding answers) archived at CQA sites to locate answers to a newly-created question Exploit wealth-of-knowledge already provided by CQA users 29 Question Answering Existing popular CQA sites Yahoo! Answers, WikiAnswers, and StackOverflow Community Question Answering System CQA-Based 30 Example. Question Answering CQA-Based Challenges for finding an answer to a new question from QA pairs archived at CQA sites 31 Misleading Answers No Answers Spam Answers SPAM Incorrect Answers Question Answering Answerer reputation CQA-Based 32 Question Answering Challenges (cont.) 300 millions posted under Yahoo! Answers since 2005: an average of 7,000 questions & 21,000 answers per hour Identifying the most suitable answer among the many available Account for the fact that questions referring to the same topic might be formulated using similar, but not the same, words CQA-Based n Matching posted questions to the best answerers who can contribute the needed information Based on the expertise/past performance of the answerers who have answered similar questions (Problem) Are the potential answerers willing to accept & answer the questions recommended to them on time? When do users tend to answer questions in a CQA system? How do users tend to choose the questions to answers in CQA? n (A solution) Analyze the answering behavior of answerers When: Analyze the overall/user-specific temporal activity patterns & identify stable daily/weekly periodicities How: Analyze factors that affect users decision, including question category, question positions, & question text 33 Question Answering CQA-Based n Applying a question-routing scheme that considers the answering, commenting & voting propensities of a group of answerers (Question) What routing strategy should be employ to ensure that a question gets answers with lasting value? (A Solution) Answerers collaborate to answer questions, who are chosen according to their compatibility, topical expertise & availability, to offer answers with high values QA process is a collaborative effort that requires inputs from different types of users User-user compatibility is essential in CQA services Evaluating topics, expertise & availability are critical in building the framework for achieving the goal of a CQA system 34 Question Answering CQA-Based n Increasing the participation of expert answerers by using a question recommendation system to proactively warn answerers the presence of suitable questions to answer (How?) Using community feedback tools, which serve as a crowd-sourced mechanism Users can vote, positively or negatively, for questions or answers, which are casted into a single score & serve as a proxy for question/answer quality (Another Solution) Using the present of text (in questions & answers) for modeling the experts & the questions. Users & questions are represented as vectors of latent features Users with expertise in similar topics are likely to answer similar questions, which can be recommended to expert users 35 Question Answering CQA-Based Corpus-based approaches Analyze text documents from diverse online sources to locate answers that satisfy the info. needs expressed in a question Overview QA SYSTEM When is the next train to Glasgow? Question 8:35, Track 9. Answer Text Corpora & RDBMS Data sources Question Extract Keywords Query Search Engine Corpus Docs Passage Extractor Answers Answer Selector Answer 36 Question Answering Corpus-based Classification: 37 Question Answering Corpus-based Challenges Factoid vs. List (of factoids) vs. Definition Open vs. Closed domain What lays blue eggs? -- one fact Name 9 cities in Europe -- multiple facts What is information retrieval? -- textual answer Identifying actual users information needs Converting to quantifiable measures Answer ranking What is apple? Magic mirror in my hand, who is the fairest in the land? Corpus-Based QA Systems n Corpus-based QA systems rely on a collection of docs, attempting to retrieve correct answers to questions posed in natural languages 38 Question Answering n Question Processing Module: Given a question Q as input, the module process, analyzes, creates a representation of the information requested in Q, and determines The question type (such as informational) based on a taxonomy of possible questions already coded into the system, e.g., Who: asking for people Where: referring to places/locations When: looking for time/occasion Why: obtaining an explanation/reason What: requesting specific information How: describing the manner that something is done The expected answer type through semantic processing of Q The question focus, which represents the main information that is required to answer Q 39 40 Sample types of questions, their corresponding answer types, and statistics from the set of TREC 8 questions Question/Answer Classification n Question Type Classification: provide constraints on what constitutes relevant data, the nature of the answer Using Support Vector Machines (SVM) to classify Q based on feature sets, e.g., text (a bag of words) or semantic (named entities) features, e.g., proper names/adjectives n Answer Type Classification: mapping question typeto answer types can be a one-to-many mapping, since question classification can be ambiguous, e.g., what n Question Focus: is defined as a word or a sequences of words indicating what info. is being asked in Q, e.g., What is the longest river in New South Wales has the focus on longest river in the question type of what Using pattern matching rules to identify the question focus 41 Question Answering n Paragraph Indexing Module (or Document Processing Module) relies on one or more IR systems to gather info. from a collection of document corpora Filter paragraphs, retaining non-stop, stemmed words Perform indexing on remaining keywords in paragraphs Access the quality of indexed (keywords in) paragraphs & order the extracted paragraphs according to how plausible they contain answers to questions (e.g., based on the question keywords in the paragraphs) 42 Question Answering n Answer Processing Module is responsible for identifying & extracting answers from paragraphs passed to it Answer Identification determines paragraphs which contain the required answer type based on named entity recognition/part-of-speech tagger to recognize answers Answer Extraction retrieves relevant words/phrases in answers to the given question Answer Correctness can be verified by the confidence in the correctness of an answer based on the lexical analysis (using WordNet?) on the correct answer type Types of answers to questions & questions are in the same domain 43 Community Based Question Answering Some complex information needs cant be answered by traditional search engines No single webpage may exist that satisfies the information needs Information may come from multiple sources Human (non-)experts in a wide range of topics form a community- based question answering (CQA) group, e.g., Yahoo! Answers CQA tries to overcome these limitations Searcher enters questions Community members answer questions 44 Example Questions 45 Community Based Question Answering Pros Users can find answers to complex or obscure questions with diverse opinions about a topic Answers are from humans, not algorithms, that can be interacted with who share common interests/problems Can search archive of previous questions/answers, e.g., Yahoo! Answers Cons Some questions never get answered Often takes time (possibly days) to get a response Answers may be wrong, spam, or misleading 46 Community Based Question Answering Yahoo! Answers, a community-driven question-and- answer site launched by Yahoo! on July 5, 2005Yahoo! 47 Question Answering Models How can we effectively search an archive of question/ answer pairs databases? Can be treated as a translation problem Translate a question into a related/similar question which likely have relevant answers Translate a question into an answer: less desirable The vocabulary mismatch problem Traditional IR models likely miss many relevant questions Many different ways to ask the same question Stopword removal and stemming do not help Solution: consider related concepts (i.e., words)the probability of replacing one word by another 48 Question Answering Models Translation-based language model (for finding related questions, then answers): translate w (in Q) from t (in A) where Q is a question A is a related question in the archive V is the vocabulary P(w | t) are the translation probability P(t | A) is the (smoothed) probability of generating t given A Anticipated problem: a good (independent) term-to-term translation might not yield a good overall translation Potential solution: matches of the original question terms are given more weight than matches of translated terms 49 Question Answering Models Enhanced translation model, which extends the translation-based language model on ranking Q: where controls the influence of the translation probability is a smoothing parameter |A| is the number of words in question A C w is count of w in the entire collection C, and |C| is the total number of word occurrence in C when 1, the model becomes more similar to the translation- based language model when 0, the model is equivalent to the original query likeli- hood model, without influence from the translation model 50 Computing Translation Probabilities Translation probabilities are learned from a parallel corpus Most often used for learning inter-language probabilities Can be used for intra-language probabilities Treat question-answer pairs as parallel corpus Translation probabilities are estimated from archived pairs (Q 1, A 1 ), (Q 2, A 2 ), , (Q N, A N ) Drawbacks Computationally expensive: sum over the entire vocabulary, which can be very large Solution: considering only a small number (e.g., 5) of (most likely) translations per question term 51 Sample Question/Answer Translations 52 Document Filtering Ad hoc retrieval Document collections are static and information needs change with time Results returned when query is entered Document filtering Document collections change with time, but (long-term) information needs are static Long term information needs represented as a profile Documents entering system that match the profile are delivered to the user via a push mechanism Must be efficient and effective (minimizes FPs and FNs) 53 Profiles Represents long-term information needs and personalizes the search experience Can be represented in different ways by including A Boolean or keyword query Sets of relevant and non-relevant documents Social tags and named entities Relational constraints Published before 1990 Price in the $10 - $25 range Actual representation usually depends on the underlying filtering model Static (filtering) or updated over time (adaptive filtering) 54 Soft filters Hard filters (must be satisfied) Document Filtering Scenarios 55 Document Stream t = 2t = 3t = 5t = 8 Profile 1 Profile 2 Profile 3 Document Stream t = 2t = 3t = 5t = 8 Profile 1 Profile 2 Profile 3 Profile 1.1 Profile 2.1 Profile 3.1 Static Filtering Adaptive Filtering Easier to process, less robust More robust, requires frequent updates Static Filtering Given a fixed profile, how can we determine if an incoming document should be delivered? Treat as an IR problem Boolean Vector space Language modeling Treat as supervised learning problem Nave Bayes Support vector machines 56 Require predefined threshold value Static Filtering with Language Models Assume a profile P consists of K relevant documents T i (1 i K) each with weight i Probability of a word w given the profile P is i is the weight (important) associated with T i f w,T i is the frequency of occurrence of w in T i is a smoothing parameter C w is count of w in the entire collection C, and |C| is the total number of word occurrence in C 57 Static Filtering with Language Models Probability of a word w given a new document D is P(w | D) = (1 - ) + KL-divergence between profile and document model is used as score If -KL(P || D) , then deliver D to P, where is some relevance threshold 58 f w,D |D||D| CwCw |C||C| True Distribution Approximation Distribution Not depended on D: ignore for ranking KL(P || D) = t V P(t | P) log P(t | D) P(t | P) Adaptive Filtering In adaptive filtering, profiles are dynamic How can profiles change (from static to dynamic)? 1. User can explicitly update the profile 2. User can provide (relevance) feedback about the documents delivered to the profile 3. Implicit user behavior can be captured and used to update the profile 59 Adaptive Filtering Models Rocchio Profiles treated as vectors Relevance-based (profiles) language models where C is the set of documents in the collection P(D i | D), the probability that D i is generated from Ds LM, which is very close to 1 if D i = D; 0, otherwise 60 Fast Filtering with Millions of Profiles Real filtering systems May have thousands or even millions of profiles Many new documents will enter the system daily How to efficiently filter in such a system? Most profiles are represented as text or a set of features Build an inverted index for the profiles Distill incoming documents as queries and run against index 61 Evaluation of Filtering Systems Definition of good depends on the purpose of the underlying filtering system Do not product ranking of documents for each profile Evaluation measures, such as and MAP, are non-relevant; precision, recall, and F-measure are computable Generic filtering evaluation measure: = 2, = 0, = -1, and = 0 are widely used 62 Summary of Filtering Models 63 Recommender Systems (RSs) n RSs are software tools providing suggestions for items to be of use to users, such as what items to buy, what music to listen to, or what online news to read n Primarily designed to evaluate the potentially overwhelming number of alternative items may offer The explosive growth & variety of information on the Web frequently lead users to make poor decisions Offer ranked lists of items by predicting the most suitable products or services based on the users preferences & constraints Often rely on recommendations provided by others in making routine, daily decisions, the collaborative-filtering technique Use various types of knowledge & data about users/items 64 Recommender Systems n Content-based Recommender Systems: Try to recommend new items similar to those a given user has liked in the past Identify the common characteristics of items being liked by user u and recommend to u new items that share these characteristics An item i, which is a text document, can be represented as a feature vector x i that contains the TF-IDF weights of the most informative keywords A profile of u, denoted profile vector x u, can be obtained from the contents of items rated by u, denoted u, and each item i rated by u, denoted r ui x u = i u r ui x i which adds the weights of x i to x u 65 a scalar value Recommender Systems n Content-based Recommender Systems: Approach: analyze the descriptions of items previously rated by a user & build a user profile (to present user interests/ preferences) based on the features of the items 66 Through ML Extract Relevant Features Matching Items +/- Feedbacks Content-Based Filtering n Advantages: User Independence: explores solely ratings provided by the user to build her own profile, but not other users (as in collaborative filtering) Transparency: recommendations can be explained by explicitly listing content features that caused an item to be recommended New Items: items not yet rated by any user can still be recommended, unlike collaborative recommenders which rely solely on other users rankings to make recommendations 67 Content-Based Filtering n Shortcomings: Limited Content Analysis: there is a natural limit in the number/types of features that can be associated with items which require domain knowledge (e.g., movies) Over-specialization: tendency to produce recommendations with a limited degree of novelty, i.e., the serendipity problem, which restricts its usefulness in applications New User: when few ratings are available (as for a new user), CBF cannot provide reliable recommendations 68 Recommender Systems n Collaborative Filtering Recommender Systems: Unlike content-based filtering approaches which use the content of items previously rated by users Collaborative filtering (CF) approaches rely on the ratings of a user, and those of other users in the system Intuitively, the rating of a user u for a new item i is likely similar to that of user v if u and v have rated other items in a similar way Likewise, u is likely to rate two items i and j in a similar fashion, if other users have given similar ratings to i & j CF overcomes the missing content problem of the content- based filtering approach through the feedback, i.e., ratings, of other users 69 Collaborative Filtering n Collaborative Filtering Recommender Systems: Instead of relying on content, which may be a bad indicator, CF are based on the quality of items evaluated by peers Unlike content-based systems, CF can recommend items with very different content, as long as other users have already shown interested for these different items Goal: identify users whose preferences are similar to those a given user has liked in the past Two general classes of CF methods: Neighborhood-based methods Model-based methods 70 Collaborative Filtering n Neighborhood-based (or heuristic-based) Filtering: User-item ratings stored in the system are directly used to predict ratings for new items, i.e., using either the user- based or item-based recommendation approach User-based: evaluates the interest of a user u for an item i using the ratings for i by other users, called neighbors, that have similar rating patterns The neighbors of u are typically users v whose ratings on the items rated by both u and v are most correlated to those of u Item-based: predicts the rating of u for an item i based on the ratings of u for items similar to i 71 Neighborhood-Based Recommendation n Example. n Eric & Lucy have very similar tastes when it comes to movies, whereas Eric and Diane have different tastes n Eric likely asks Lucy the opinion on the movie Titanic and discards the opinion of Diane 72 Collaborative Filtering n User-based Rating Prediction: Predicts the rating r ui of a user u for a new item i using the ratings given to i by users most similar to u, called nearest-neighbors Given the k-nearest-neighbor of u who have rated item i, denoted N i (u), the rating of r ui can be estimated as r ui = If the neighbors of u can have different levels of similarity with respect to u, denoted w uv, the predicted rating is r ui = 73 1 r vi |N i (u)| vNi(u)vNi(u) w uv r vi vNi(u)vNi(u) vNi(u)vNi(u) |w uv | Collaborative Filtering n Item-based Rating Prediction: While user-based methods rely on the opinion of like-minded users, i.e., similar users, to predict a rating, item-based approaches look at ratings given to similar items Example. Instead of consulting with his peers, Eric considers the ratings on the movies he (& others) has (have) seen Let N u (i) be the set of items rated by user u most similar to item i, the predicted rating of u for i is r ui = 74 w ij r uj jNu(i)jNu(i) jNu(i)jNu(i) |w ij | Collaborative Filtering n Advantages of Neighborhood-based Filtering: Simplicity: the methods are intuitive & relatively simple to implement (w/ only the no. of neighbors requires tuning) Justifiability: the methods provide a concise & intuitive justification for the computed predictions Efficiency: the methods require no costly training phases & storing nearest neighbors of a user requires very little memory. Thus, it is scalable to millions of users & items Stability: the methods are not significantly affected by the constant addition of users, items, and ratings in a large commercial applications & do not require retraining 75 Collaborative Filtering Static and adaptive filtering are not social tasks; profiles (and their users) are assumed to be independent of each other Similar users are likely to have similar preferences or profiles Collaborative filtering exploits relationships between users profiles to improve how items (documents) are matched to users (profiles) If A is similar to B and A judged a document D is relevant, then it is likely that D is also relevant to B Often used as a component of recommender system 76 Collaborative Filtering According to [Ma 09], there are two widely-used types of methods for collaborative filtering Neighborhood-based methods Include user-based approaches, which predict the ratings of active users based on the computed information of items similar to those chosen by the active user. Suffer from data sparsity and scalability problems Model-based methods use the observed user-item ratings to train a compact model that explains the given data so that ratings can be predicted 77 [Ma 09] Ma et al. Semi-nonnegative Matrix Factorization with Global Statistical Consistency for Collaborative Filtering. In Proc. of ACM Conf. on Information and Knowledge Management Rating using User Clusters Clustering can be used to find groups of similar users Measure user/user similarity using rating correlation: Use average rating of other users within the same cluster to rate unseen items 78 Collaborative Filtering Example. Predicts the missing values in the user-item matrix [Ma 09] 79 User-item Matrix Predicted User- item Matrix Recommender Systems n Evaluating Recommendation Systems Recommendation systems have a variety of properties, such as accuracy and robustness, that may affect user experience Prediction Accuracy: Predict user opinions (e.g., ratings) over items or the probability of usage (e.g., purchasing ) Basic assumption: a recommender system that provides more accurate predictions will be preferred by the user Measuring prediction accuracy in a user study measures the accuracy given a recommendation Commonly used rating accuracy measures: Root Mean Squared Error (RMSE) or Mean Absolute Error (MAE) 80 Recommender Systems n Evaluating Recommendation Systems (Continued) Prediction Accuracy: Usage Prediction Measures (recommend items for users to use): Precision, Recall (True Positive Rate), and False Positive Rate 4 possible outcomes for the recommended & hidden items Precision = #TP / (#TP + #FP) Recall = #TP / (#TP + #FN) False Positive Rate = #FP / (#FP + #TN) 81 RecommendedNot Recommended UsedTrue-Positive (TP)False-Negative (FN) Not UsedFalse-Positive (FP)True-Negative (TN) Recommender Systems n Novelty Recommendations Recommending items that the user did not know about, but is not surprising, e.g., a new movie by the same director An offline evaluation strategy Split the data set on time by hiding all the user ratings that occurred after a specific point in time (ST) Hide ratings that occurred prior to ST When recommending, the system is rewarded for each item that was recommended & rated after ST, but is punished for each item that was recommended prior to ST 82 Recommender Systems n Serendipity Recommendations Is a measure of how surprising the successful recommend- ations are E.g., following a successful movie recommendation, the user learns of a new actor that (s)he likes n Diversity Recommendations Is generally defined as the opposite of similarity in which suggesting similar items may not be useful to the user The most explored method for measuring diversity uses item-item similarity, typically based on item content In recommenders that assist in information search, more diverse recommendations will result in shorter search interactions 83